def forward(self, x):
if self.normalize:
x = normalize_imagenet(x)
f3,f2,f1 = self.features(x)
# f3: 512 f2: 256 * 14 * 14 f1: 128 * 28 * 28
# full kmax pooling
f1 = f1.detach()
f1 = kmax_pooling(f1,1,2)
f1 = f1.view(f1.size(0), -1)
f1 = self.f1_fc(f1)
f2 = f2.detach()
f2 = kmax_pooling(f2,1,8)
f2 = f2.view(f2.size(0), -1)
f2 = self.f2_fc(f2)
f3 = self.fc3(f3)
f2 = self.fc2(f2)
f1 = self.fc1(f1)
if self.batch_norm:
f3 = self.f3_bn(f3)
f2 = self.f2_bn(f2)
f1 = self.f1_bn(f1)
return f3, f2, f1
def forward(self, x):
batch_size = x.size(0)
net = normalize_imagenet(x)
net = self.convnet(net)
net = net.view(batch_size, 256 * 3 * 3)
out = self.fc_out(net)
return out
def call(self, x, training=False):
if self.normalize:
x = normalize_imagenet(x)
x = self.conv1(x)
x = self.bn1(x, training=training)
x = tf.nn.relu(x)
x = self.pool1(x)
x = self.layer1(x, training=training)
x = self.layer2(x, training=training)
x = self.layer3(x, training=training)
x = self.layer4(x, training=training)
x = self.avgpool(x)
out = self.fc(x)
return out
def forward_local(self, data, pts):
assert self.local
world_mat = data['world_mat']
camera_mat = data['camera_mat']
x = data[None]
pts = transform_points(pts, world_mat)
points_img = project_to_camera(pts, camera_mat)
points_img = points_img.unsqueeze(1)
local_feat_maps = []
if self.normalize:
x = normalize_imagenet(x)
x = self.features.conv1(x)
x = self.features.bn1(x)
x = self.features.relu(x)
x = self.features.maxpool(x) # 64 * 112 * 112
x = self.features.layer1(x)
local_feat_maps.append(x) # 64 * 56 * 56
x = self.features.layer2(x)
local_feat_maps.append(x) # 128 * 28 * 28
x = self.features.layer3(x)
local_feat_maps.append(x) # 256 * 14 * 14
x = self.features.layer4(x)
local_feat_maps.append(x) # 512 * 7 * 7
x = self.features.avgpool(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
# get local feats
local_feats = []
for f in local_feat_maps:
#f = f.detach()
f = F.grid_sample(f, points_img, mode='nearest')
f = f.squeeze(2)
local_feats.append(f)
local_feats = torch.cat(local_feats, dim=1)
local_feats = local_feats.transpose(1, 2) # batch * n_pts * f_dim
local_feats = self.local_fc(local_feats)
# x: B * c_dim
# local: feats B * n_pts * c_dim
return x, local_feats
def forward(self, x, h, u, time):
if self.normalize:
x = normalize_imagenet(x)
net = self.features(x)
fc = self.fc(net)
rect = self.leaky_relu(fc)
t_x_s_update = self.t_x_s_update(rect, h, time)
t_x_s_reset = self.t_x_s_reset(rect, h, time)
update_gate = self.sigmoid(t_x_s_update)
complement_update_gate = 1 - update_gate
reset_gate = self.sigmoid(t_x_s_reset)
rs = reset_gate * h
t_x_rs = self.t_x_rs(rect, rs, time)
tanh_t_x_rs = self.tanh(t_x_rs)
gru_out = update_gate * h + complement_update_gate * tanh_t_x_rs
return gru_out, update_gate
def forward(self, x):
if self.normalize:
x = normalize_imagenet(x)
net = self.features(x)
out = self.fc(net)
return out
def call(self, x, training=False):
if self.normalize:
x = normalize_imagenet(x)
net = self.features(x, training=training)
out = self.fc(net)
return out
def encode_first_step(self, x):
if self.normalize:
x = normalize_imagenet(x)
f3,fs2,fs1 = self.features.calc_feature_maps(x)
return f3, fs2, fs1
def encode_first_step(self, x):
if self.normalize:
x = normalize_imagenet(x)
f3,f2,f1 = self.features(x)
return f3, f2, f1
